Advanced Deep Learning Approaches for Automated Diabetic Retinopathy Detection and Severity Classification: A Multi-Model Review

Abstract

Diabetic Retinopathy (DR) is one of the main causes of preventable blindness worldwide, especially in people of working age. Manual assessment of retinal fundus images is labor-intensive and frequently subject to variability, particularly in the early phases of the disease. Advancements in deep learning (DL) methodologies have facilitated automated, high-precision detection and grading of Diabetic Retinopathy severity, providing scalable frameworks for clinical screening applications. This review consolidates pivotal research employing hybrid deep learning (DL) architectures, such as Google Net and ResNet augmented with adaptive particle swarm optimization (APSO), in conjunction with conventional machine learning classifiers like Support Vector Machines, Random Forests, and Decision Trees. Furthermore, the surveyed literature underscores the emergence of advanced paradigms—including supervised, self-supervised, and transformer-based frameworks—and explores the integration of federated learning and generative adversarial networks (GANs) to enhance model resilience and generalizability. The review articulates prospective avenues, including the integration of multi-modal data sources and the development of resource-efficient architectures to facilitate real-world clinical implementation.